By 
Shreyas M V · Published · 13 min read
Ten AI Startup Homepages and the Claims They Can Prove
Of ten widely known AI company homepages read on 17 October 2026, most put their strongest claims (“best”, “#1”, “most accurate”) in the page text with no source, date or method beside them. Figures that did carry support were usually credited to a named customer or analyst, as on the Writer, Glean and Cursor pages, not shown with the company’s own method. That describes what the page text shows and says nothing about whether the products deliver. This page gives the four-level rubric, the scored findings for each company and the steps to run the same check on your own homepage.
What do ten AI startup homepages show beside their claims?
They show far more claims than proof: across the ten pages, superlatives and the company's own numbers usually had nothing beside them in the page text, and the best-supported figures came from named third parties. The points below are the short version.
- Superlatives such as “best”, “leading”, “#1” and “most accurate” were the most common claim type, and they sat almost always at level 3 or level 4 of the rubric.
- Where a number was backed, it was usually backed by a named customer or analyst (Writer, Glean, Cursor), not by the company's own method.
- Stated methods and dates beside numbers were rare on all ten pages.
- Three pages gave two wordings or figures for the same thing on the same page: ElevenLabs, Synthesia and Glean.
- None of this says whether any product works. It records what a reader can check without leaving the page.
How were the ten homepages chosen and read?
The ten are widely known AI companies whose homepages could be fetched as text: Cursor, ElevenLabs, Replit, Glean, Harvey, Mistral, Jasper, Synthesia, Writer and Runway. Each homepage's text was fetched on 17 October 2026 and read for its wording and for what sits beside each claim.
Two companies were tried and left out for access reasons alone: Perplexity returned no readable text and Lovable returned an HTTP 403 error. The selection is not a ranking of the best or worst homepages, and it is not a sample designed to represent AI startups.
The limits matter. This is a reading of page text. Screenshots, video, images, animations, anything behind a click and anything on a linked page were not read.
So “not in the fetched text” means that no supporting source, date, method or named attribution appeared next to the claim. It does not mean the company lacks proof elsewhere, because the linked case studies, research pages and trust centres were not opened. The readings were made by the operator of this site on one day.
What is the four-level evidence check?
The evidence check sorts every claim on a page into one of four levels by what sits beside it in the page text: a source, a named person, only a link, or nothing.
The levels were set before the claims were scored and applied to each one. They measure how visible the proof is, not whether the claim is true. A level 4 claim can be accurate, and a level 1 claim can still be a narrow one.
| Level | What qualifies | Example from the ten | What it tells a reader |
|---|---|---|---|
| Level 1: Sourced on the page | A named source, a date or a method sits beside the claim, in the page text. | Glean names a Gartner report on its Emerging Market Quadrant for No-Code Agent Builders and links it. That is a recognition, not a measured result. | The reader can check the claim without leaving the page. |
| Level 2: Attributed | A named person, with title and company, stands behind the claim. | Writer shows 70% time savings on derivative content, credited to Lauren Boyman, CMO at KPMG. | Who said it is known. How it was measured is not. |
| Level 3: Linked only | A link or logo suggests proof exists elsewhere, but the proof is not in the page text. | Jasper ties 10,000+ hours saved annually to a Cushman & Wakefield case study link. | Proof may exist. This reading did not open linked pages. |
| Level 4: Bare claim | No source, date, method or attribution beside the claim in the page text. | Replit states a 65% lower average cost than the previous version of Max Mode, with no benchmark, method or date. | The reader has nothing on the page to check. |
Two further notes were recorded for every page: whether a claim is a superlative or a number, and whether it differs from another figure on the same page. Both come up below.
How do the ten homepages compare at a glance?
Every page had at least one claim the reading placed at level 4, and no page was free of linked-only or attributed support. The table gives each page's best-supported item, the level it earned and one typical bare claim.
| Company | Strongest supported item | Level | A typical level 4 claim |
|---|---|---|---|
| Cursor | Named testimonials with titles and figures, such as Diana Hu of Y Combinator on adoption passing 80% | 2 | “Use the best model for every task” |
| ElevenLabs | A “$1B In-Kind Donation” headline link, and a linked “most accurate ASR model” claim | 3 | “98% accuracy” for Scribe and “75ms latency” |
| Replit | Named customer quotes from Gusto, Talkdesk and Zillow, all opinion with no metrics | 2 | “65% lower average cost” than the previous Max Mode |
| Glean | A Gartner mention with the report linked, and customer cards that link to stories | 1 and 3 | A stats block: “110 hours saved per user/year”, “99.9%+ uptime” |
| Harvey | SOC 2 II and ISO certifications with links; named quotes with no metrics | 3 and 2 | “25+” average hours saved per month and “top law firms” |
| Mistral | Customer logos with quotes attributed by job title only | 3 | “#1 in Sovereign AI” |
| Jasper | Case study and webinar links behind two customer figures | 3 | “100+ specialized AI agents” and “up to 10x faster” |
| Synthesia | Named customer results, such as Moody’s going from 4 hours to 30 minutes | 2 and 3 | “#1 AI Video Platform for Business” and “Over 2,000,000 users” |
| Writer | Three figures each credited to a named person with a title | 2 | “10x” and “campaign-ready assets in minutes, not weeks” |
| Runway | A Gen-4.5 claim that links to a research page | 3 | “Used by 60m+ creatives” and “The World’s Best Models” |
The table is a summary, not a score. A homepage with a level 3 best item is not worse than one with a level 2 item, because the readings did not weigh how many claims each page made.
What did the first five homepages show beside their claims?
Cursor and Glean tied numbers to named third parties, while ElevenLabs, Replit and Harvey left most numeric claims bare in the page text. Each entry below states what the text showed.
Cursor
Cursor's homepage puts its numbers in other people's mouths. Testimonials name Diana Hu of Y Combinator (adoption “went from single digits to over 80%”), Jensen Huang of NVIDIA (“some 40,000” engineers), Patrick Collison of Stripe and Greg Brockman of OpenAI.
The figures carry no method or date. “Trusted by over half of the Fortune 500” has only an enterprise link beside it, and “Use the best model for every task” has nothing. The footer lists SOC 2, ISO27001, ISO42001 and AIUC-1 with a link to a security page that was not read.
ElevenLabs
On the ElevenLabs homepage, most numeric claims had no method beside them: “98% accuracy” for Scribe and “75ms latency for conversational usecases” for Eleven Flash. “Independently rated the leading Text to Speech models” says “independently” but names no rater, link or method in the text.
Timeline entries dated November 2025 and January 2026 repeat the “most accurate” wording with no benchmark. They are dated, which is more than most superlatives get, but a date alone does not make a claim checkable. The page's conflicting figures are covered below.
Replit
Replit attributes only opinion. Named quotes come from Alex Meyers of Gusto (“10x easier”), Shauna Geraghty of Talkdesk (“in hours rather than weeks”) and Doug Rodermund of Zillow (“unmatched”), among others, and none carries a metric.
The numeric product claims are bare. “65% lower average cost than the previous version of Max Mode” names no benchmark, method or date, and “100+ integrations” lists examples such as OpenAI, Stripe and Google Workspace without a count source. Plan terms such as the 28-day rollback are pricing terms, not performance claims, and were set aside.
Glean
Glean splits cleanly in two. The platform-level stats block has no date, period or method: “40+ Unique LLMs”, “110 hours saved per user/year”, “93% enterprise adoption in < 2 years”, “3+ billion” searches, “99.9%+ uptime” with no period, and a live counter of token savings with no baseline.
The customer side is better supported. Cards for Booking.com, Zillow, TIME, Ericsson and GCash link to stories, the Zillow card adds a named quote from Toby Roberts that repeats the 80% adoption figure, and the page names a Gartner report on its Emerging Market Quadrant for No-Code Agent Builders with the report linked. “Best-in-class data governance” sits beside ISO 42001, HIPAA and SOC 2 Type II badges, but no comparison supports “best-in-class”.
Harvey
Harvey pairs a confident subhead with a bare stats block. “The world's top law firms” has no ranking or source, and the figures for hours saved per month, professionals, legal organizations, countries and AmLaw 100 firms carry no date, period or sample.
Named quotes from Rich Robbins of Reed Smith and Robert Clark of Dentsu are attributed but contain no metrics. The security section lists SOC 2 II, ISO 27001, ISO 27701 and ISO 42001 with links, and a banner about a $550M raise at a $15.5B valuation links to a blog post. “The most powerful Harvey yet” compares the product with its own earlier version and shows no measure.
What did the other five homepages show beside their claims?
Writer attached a named person and title to each of its three headline figures, Synthesia and Jasper linked customer results to stories or case studies, and Mistral and Runway leaned on superlatives, logos and links to research.
Mistral
Mistral headlines “#1 in Sovereign AI” with no source, ranking, date or definition of the category beside it. “Europe's three largest banks trust Mistral” and “5 of the world's largest energy companies trust Mistral” are shown with logos and quotes attributed by job title only, with no person named.
“NATO and 20+ national government entities trust Mistral” appeared with no NATO logo or quote in the text and no list or date for the “20+”. Outcome lines such as ASML “identifying sub-system errors 70% faster” come with a logo and no baseline, method or date. No benchmark with a method and no dated figure appeared in the page text, and “frontier open models” links to a research page that was not read.
Jasper
Jasper links its two most specific figures to proof elsewhere: 10,000+ hours saved annually points to a Cushman & Wakefield case study, and 7,500 product descriptions in 24 hours points to an Adidas webinar page. Neither has a date or method in the page text.
Other claims were bare, including “100+ specialized AI agents”, “60% of SEO now automated with Jasper” (a logo sits nearby but no customer is named) and the FAQ's “up to 10x faster” and “up to 50% lower cost”. The page's structured data states a 4.8 rating from 1200 reviews, yet the visible text shows no rating and names no platform or date.
Synthesia
Synthesia uses “#1 AI Video Platform for Business” in its page title and footer, and “The #1 rated AI Video Platform” elsewhere, with no ranking source. “Trusted by 90% of the Fortune 100” sits over a logo strip with no dated figure, and “Over 2,000,000 users” and “Trusted by 50,000+ teams” have nothing beside them.
Customer results are named at company level: Moody's going from 4 hours to 30 minutes, Mondelez at 100 hours of translation in 10 minutes, Linde over 50% faster, OpenText 50% faster and Five Below saving $56K. The page cites “2,000+ 5-star reviews on G2” with one named review and the G2 logo, and no link in the text.
Writer
Writer had the most consistently attributed numbers of the ten. Each headline figure has a named person and title behind it: 70% time savings on derivative content from Lauren Boyman, CMO at KPMG; 57% faster time to market from Adam Uhrich at Vanguard; and “north of 85%” savings from Matt Harker at Clorox.
No dates or methods appear beside them, and the Clorox figure has no case study link in the text. Phrases such as “campaign-ready assets in minutes, not weeks” and “10x” are bare, and “Five-year track record of AI research and LLM innovation” links to the research lab.
Runway
Runway makes its biggest claims in superlatives: “The world's best video model” for Gen-4.5 and “state-of-the-art General World Model” for GWM-1. The Gen-4.5 line links to a research page, but no benchmark or method appears in the page text.
“Used by 60m+ creatives around the world” and “We partner with the world's leading organizations” are bare, and the second names no partner in the text.
Which kinds of homepage claim were backed, and which were not?
Claims that could be handed to a third party were backed more often than claims the company has to make about itself. The table groups the claims by type and shows the level the reading placed them at.
| Claim type | Where it appeared | Usual level |
|---|---|---|
| Superlatives: “best”, “leading”, “#1”, “most accurate”, “state-of-the-art” | ElevenLabs, Mistral, Synthesia, Runway, Cursor, Replit | Almost always level 3 or 4 |
| The company’s own scale: users, teams, connectors, countries | Runway (60m+ creatives), Synthesia (2,000,000 users), Glean (275+ connectors), Harvey (85+ countries) | Level 4 |
| Speed, cost or time-saved figures from the company | Replit (65% lower cost), ElevenLabs (75ms), Jasper (up to 50% lower cost), Glean (110 hours saved) | Level 4 |
| Customer outcomes | Writer (KPMG, Vanguard, Clorox), Synthesia (Moody’s), Glean (Zillow), Mistral (ASML, BNP Paribas) | Level 2 where a person is named, level 3 where only a logo or link is shown |
| Certifications and analyst recognition | Cursor, Glean and Harvey (security badges), Glean (Gartner) | Level 3, and level 1 for the Gartner mention in part |
Counting from the first table, one of the ten pages, Glean, had any item placed even partly at level 1, and that item was a recognition, not a measured result. Writer, Glean and Cursor attached figures to named people or a named analyst. Harvey's named quotes carried no numbers.
This is analysis, and a reasonable one: a customer will put a name to a result, but a company rarely publishes the method behind its own speed or accuracy claim on a homepage. The rubric cannot tell whether that method exists elsewhere.
Where did a homepage disagree with itself?
Three of the ten pages gave two wordings or figures for the same thing: ElevenLabs, Synthesia and Glean. The table lists them, with the explanation the page text does not supply.
| Page | The two figures or wordings | What the text leaves open |
|---|---|---|
| ElevenLabs | Voices: “5,000+” in the meta description, “10,000+” in the voices section | The two sit in different places on one page. |
| ElevenLabs | Languages: “90+” in one place, “29+” for the Text to Speech API models | The second figure is scoped to the API models, so the numbers may count different things. The text does not say. |
| Synthesia | Languages: “140+” repeated, “130+” in the AI Dubbing section | The dubbing section may count a narrower set. The text does not say. |
| Glean | Booking.com: “1st AI platform adopted worldwide” in the card headline, “first AI platform adopted company-wide” in the body | A headline and its own body text word the same milestone differently. |
These may all be explainable. A different scope, a different counting rule or an older line of copy would each produce the same pattern. The point for a founder is that a reader who spots the gap has no way to resolve it from the page.
What about numbers inside product demos and offers without a year?
Numbers inside product demos look like data, but the page text does not mark them as examples. Cursor's demo panel shows “Worked for 14m 22s”, which is illustrative interface text, not a sourced figure.
ElevenLabs shows an ElevenAgents graphic with a resolution rate of 83.4%, with 87.37% for V1 and 61.71% for V2 on 17 and 24 August. The graphic shows no customer, source or method, so a reader cannot tell a real result from a mock-up. Demo panels are a normal part of a product page, and labelling one as an example removes the question.
Dating is the cheap fix for the other gap. Runway offers enterprises two months of usage free and says to sign by 10/31, with linked terms but no year in the text. A reader on 17 October 2026 can guess the year, a reader on a cached copy cannot.
What is documented and what is our analysis?
The documented list is what the page text showed on 17 October 2026, and the analysis list is Kanso's interpretation of it. Keeping them apart lets you disagree with the second without doubting the first.
Documented (observed in the page text)
- The exact wording and the neighbouring sources, names, dates and links for each claim quoted above.
- The three same-page differences at ElevenLabs, Synthesia and Glean.
- That Jasper's structured data states a 4.8 rating from 1200 reviews while its visible text does not.
- That Perplexity returned no readable text and Lovable returned an HTTP 403, so neither was included.
Analysis (inferred from what was observed)
- Superlatives were the most common claim type and were almost always level 3 or 4.
- Backed numbers usually came from a named customer or analyst, not the company's own method.
- Stated methods and dates beside numbers were rare across all ten.
- A clearer scope or date beside a figure would settle the same-page differences.
None of it covers whether the products deliver what they claim, and a single day's reading of ten pages cannot support a claim about all AI startups. Companies named here are welcome to ask for a correction, and a corrected reading will be dated and noted on this page.
How do the pages that already rank treat claims?
The three ranking pages read for this article judge design and persuasion, and none of the portions read tests whether the claims on an example are sourced. Ebaq Design's seven AI startup examples concentrate on layout, motion and visual identity, Landingi's five startup examples list takeaways such as layout, headline and trust signals, and Unbounce's 40 examples explain what makes each page persuasive.
Their own statistics are a fair test of the rubric. Unbounce states that landing pages with video can increase conversions by 86%, and the text read shows no study, date or method beside it, which is level 4. It also cites its own Conversion Benchmark Report for media and entertainment pages under 350 words converting better, which names a source and so sits at level 1 on the name alone, though the text read does not show a date.
Landingi credits Waseem Bashir of Apexure with an estimate of $113 as the typical cost of designing a landing page with tools, which is level 2: a named person, no method. Unbounce's line that “recent reports show” Netflix at over 300 million subscribers names no report or date, so it is level 4. The links beside these lines were not visible in the text read.
These pages are the better read when you want visual inspiration, layout patterns or a build tool, and each of the three publishers sells something in that area: Ebaq Design designs websites, Landingi sells a page builder and Unbounce sells one too. This page is the better read when the question is what a reader can verify. For a related look at how one company turns its documentation into marketing, with the same attention to what is sourced, see the Stripe documentation teardown.
How can you run the evidence check on your own homepage?
You can run it in about an hour with a text copy of the page and a notes file, following the seven steps below. The method mirrors how the ten pages were read, so your results can be compared with the table above.
- Capture the text and date it. Save the page text as it stands today. Homepages change often, so the date belongs in your notes.
- List every claim. Include each number, each ranking word (“best”, “#1”, “leading”) and each “trusted by” line.
- Look beside each claim. Find any source, date, method, named person with title, link or logo that sits next to it.
- Assign a level. Use levels 1 to 4 from the table above, and note when a claim is a superlative or a number.
- Search for conflicts. Search the page for each figure's variants, such as two language counts, and note any pair that differs.
- Decide per level 4 claim. Put the source, date or method next to it, reword it into something you can show, or cut it.
- Re-date the page. Write the capture date next to any figure that can change, and repeat the check when the copy changes.
One point of analysis from the rubric: a link to a case study moves a claim from level 4 to level 3, not to level 1. To reach level 1, the source, date or method has to be in the page text itself.
Kanso's own homepage should be held to the same rubric. It was not scored for this page, and nothing here claims it would score better than the ten above.
If you track rivals each week, add their homepage claims to that log. The competitor check in the weekly growth checklist already asks you to note changes to a competitor's homepage headline, pricing page and plan names.
AI startup landing page examples: frequently asked questions
What makes a landing page effective?
A landing page is effective when it serves a single clear goal, uses social proof and a strong headline to build trust, and guides the visitor to one prominent call to action, according to Unbounce’s list of landing page examples. This teardown adds one test to that advice: whether the proof is visible on the page, with a source, a date or a method, or only implied by a logo or a link.
What are the key elements of an effective startup landing page?
Landingi lists a clear and intuitive layout, a unique value proposition, benefit-oriented language, high-quality visuals, strong calls to action with simple forms, and social proof. The evidence check in this article applies to the last item: social proof counts for more when a reader can see who said it, when, and how it was measured.
What is a startup landing page?
Landingi defines a startup landing page as a dedicated page built to convert visitors into leads, customers or subscribers, usually around one goal such as a sign-up, a launch or a download. The ten pages read here are full homepages of AI companies, which carry broader goals than a single campaign page, though the same claim-and-proof question applies to both.
How can I optimize my startup landing page for higher conversion rates?
Landingi advises setting achievable conversion goals from your industry’s average conversion rate, then running A/B tests on headlines, calls to action and designs and adjusting from analytics. This article does not measure conversion and has no data on whether sourced claims convert better than bare ones, so treat the rubric as a credibility check, not a conversion tactic.
Are AI website builders good enough for startup landing pages now?
This article did not test any website builder, so it cannot rate them. What it does show is that the proof behind a homepage claim is content, not tooling: whichever tool builds the page, a number or a ranking still needs a source, a date or a method beside it to be checkable.
Which AI startup homepage claims were backed by evidence on the page?
Few were backed on the page itself. Across the ten homepages read on 17 October 2026, figures attached to a named person with a title were the most consistent form of support, as on Writer’s page, and a Gartner mention on Glean’s page was the only item placed even partly at level 1. Stated methods and dates beside numbers were rare.
Does a claim with no source beside it mean the company has no evidence?
No. “Not in the fetched text” means no source, date, method or named attribution appeared next to the claim in the page text, and the linked case studies, research pages and trust centres were not opened. A bare claim is a gap in what the page shows, not a finding about the product.
Do superlatives like “best” or “#1” need a source?
Under this rubric they do, because “#1” and “best” imply a comparison that a reader can only check if the ranking, the date and the category are named. Mistral’s “#1 in Sovereign AI” and Synthesia’s “#1 AI Video Platform for Business” both appeared in the page text with no ranking source beside them. Marketing convention often allows such claims, so the rubric records the gap and does not call them false.
Why would a homepage show two different numbers for the same thing?
The causes the page text cannot rule out are different scopes, stale copy and different counting rules. ElevenLabs gave “90+” and “29+” languages, but the second was for its Text to Speech API models, and Synthesia gave “140+” and “130+”, the second in its dubbing section. Stating the scope beside each figure removes the doubt.
How do I check my own homepage claims?
List every number and every superlative on the page, look beside each for a source, a date, a method or a named person, and give it a level from 1 to 4. Then add proof next to each level 4 claim, reword it to something you can show, or remove it. The steps are set out in the section on running the check yourself.
Sources
- Cursor homepage, text read 17 October 2026, Cursor
- ElevenLabs homepage, text read 17 October 2026, ElevenLabs
- Replit homepage, text read 17 October 2026, Replit
- Glean homepage, text read 17 October 2026, Glean
- Harvey homepage, text read 17 October 2026, Harvey
- Mistral homepage, text read 17 October 2026, Mistral AI
- Jasper homepage, text read 17 October 2026, Jasper
- Synthesia homepage, text read 17 October 2026, Synthesia
- Writer homepage, text read 17 October 2026, Writer
- Runway homepage, text read 17 October 2026, Runway
- 7 Best AI Startup Landing Page Examples in 2026, Ebaq Design
- Startup Landing Page: Definition, How to Create & 5 Examples, Landingi
- 40 Best Landing Page Examples of 2026 (For Your Swipe File), Unbounce
Note: This is a reading of homepage text on 17 October 2026, not a test of any product. Homepages change often, so a link opens the live page, which may no longer match what is described here. Companies named here are neutral subjects of the same check, and Kanso has no affiliation with any of them.
Teardowns